Gavin Baker Goes All-In on AI Infrastructure: July Selloff Completely Diverges from Fundamentals, NVIDIA Sits at Lowest Forward P/E in a Decade

Compiled & Edited by: Odaily TechFlow

Guest: Gavin Baker, Founder & Chief Investment Officer, Atreides Management

Host: Patrick O'Shaughnessy, Host of Invest Like The Best, CEO of Positive Sum

Podcast Source: Invest Like The Best

Original Title: The AI Selloff Doesn't Match the Data | Top AI Investor Explains

Air Date: August 4, 2026

Disclosure: Atreides Management, founded by this episode's guest Gavin Baker, manages approximately $5 billion in assets, with significant holdings in NVIDIA (over 20 years), Astera Labs (roughly 10% of the fund), Micron, Cerebras, and other AI infrastructure names discussed in this article. His views have a direct financial interest tied to his holdings.

That said, he is indeed a top-tier technology investor in the United States, having previously served as a Managing Director at Fidelity, where he managed a multi-billion dollar technology fund. As an early and long-term investor in NVIDIA, Tesla, and SpaceX, Gavin brings over two decades of experience investing in semiconductors and cutting-edge technology. Amidst the AI浪潮, he is renowned across Silicon Valley and Wall Street for his deep insights into the chip supply chain, compute bottlenecks, and the commercialization pathways of AI.

Key Takeaways

Gavin Baker is the founder of Atreides Management and has been a persistent heavyweight investor in NVIDIA for the past 20 years, establishing himself as one of the most steadfast bulls in the AI infrastructure space. This episode was recorded following the July selloff in AI stocks, during which he returned from Silicon Valley specifically to "stress test" all of his assumptions. His core conclusion: July's selloff was completely disconnected from fundamental data – every quantitative indicator was accelerating while stock prices were in freefall.

Baker describes July as "compressing 2022 into a single month." A cohort of AI stocks fell 40% to 60% from their highs, yet GPU spot prices rose rather than fell (up 50%-60% within six months), hyperscaler operating cash flow accelerated from 28% to 35% (after adjusting for one-time items), and token output continued to grow. He believes the market made three misjudgments: interpreting Meta's compute rental as oversupply (actually mirroring SpaceX's strategy of selling compute at high prices), interpreting the surge in open-source models as shrinking demand (actually just transferring profits from the frontier model layer to the infrastructure layer), and interpreting widening CDS spreads as a credit crisis signal (actually just banks hedging on commitments). The only thing that genuinely makes him nervous is rising real yields and tightening credit markets, but his calculations suggest that if installed compute is repriced at current spot rates, hyperscalers could largely fund their AI infrastructure investments through operating cash flow, with minimal reliance on debt.

Highlight Reel

1. On the Divergence Between the Selloff and Fundamentals

"I spent two months in Silicon Valley and didn't hear a single negative quantitative indicator. GPU availability, GPU rental prices, DRAM spot prices, token growth – every single metric is accelerating."

"NVIDIA is currently trading at its lowest forward P/E ratio in the past decade. The market is 100% convinced that AI companies are massively overstating their earnings."

"OpenAI has accelerated, Anthropic is growing strongly, and open-source models are exploding. But the public markets can't see Anthropic's or OpenAI's data, creating an information asymmetry."

2. On Open-Source Models and Compute Demand

"A token is a token. Whether it's generated by a frontier model or an open-source model, it takes the same amount of compute, memory, and electricity to produce."

"What open-source models take away from frontier models is profit margin, not compute demand. It turns a 90% gross margin token into a 30% gross margin token, but the underlying GPU hours consumed are actually higher."

"Jensen is the world's biggest open-source advocate. If it were bad for his business, would he really champion it as a signature issue?"

3. On Claude's Impact on the Market

"Claude is essentially the Walter Cronkite of the stock market. Everyone feeds the news into Claude, then trades based on Claude's interpretation. Claude is smart, but it isn't always right."

"Someone posted a chart of Japanese capacitor stocks, saying 'we just ran the entire capacitor cycle in 6 weeks.' Fundamentals haven't even arrived, yet the stock price has already surged and crashed."

4. On the Game Theory of the Memory Market

"Suppose in 2027 or 2028 you want to tear up an LTA (Long-Term Agreement) to get a lower price. But if leverage shifts back to memory manufacturers within the next few years, you're out. You could destroy your entire business."

"What NVIDIA is doing, essentially, is providing a credit wrapper to GPU buyers, plus a revenue share above a floor price. This effectively allows them to rapidly build a massive cloud business through royalties."

5. On SpaceX and Orbital Compute

"SpaceX has accessed more compute faster and at lower prices than anyone else over the past three years. Now there are reports they're going to deploy 8 gigawatts of compute. I never bet against Elon, but that is a truly staggering number."

"Orbital compute becomes more real every day. Benchmark invested in StarCloud, an orbital compute company that doesn't have SpaceX's internal launch cost advantages. If Benchmark folks thought it wasn't viable, they wouldn't have invested."

6. On Regulation and the PR Crisis

"Data centers are the best thing I've ever seen for blue-collar wages. But the Democratic party, which is supposed to represent blue-collar workers, is pushing their jobs away."

"Someone in an academic book overestimated data center water usage by a factor of ten thousand. The author has admitted the mistake multiple times, but it's like the spinach iron story: a lie travels halfway around the world while the truth is still putting on its shoes."

Chapter One: "Compressing 2022 into One Month"

Patrick O'Shaughnessy: What happened this month?

Gavin Baker: I would describe July as "compressing 2022 into a single month." There are certainly some negative fundamental factors to discuss, but overall, the fundamental balance is improving significantly. A cohort of AI stocks fell 40% to 60% from their highs – a sheer vertical drop. Let me ask you, you spent the summer in Silicon Valley – did you hear even one negative quantitative indicator about AI?

Patrick O'Shaughnessy: A signal of deceleration?

No. In fact, every indicator is accelerating. Whether it's GPU availability, GPU rental prices, this month's DRAM spot prices, or token growth – all accelerating.

I think a large part of the problem is that the public markets have no visibility into Anthropic and OpenAI. Then you have these open-source inference clouds – Fireworks, Baseten, Modal, Together – they're commercializing inference. Open-source models have accelerated massively because of GLM 5.2, Kimi K3, and Nemotron is making steady progress. OpenAI is accelerating, Anthropic is growing strongly, and almost certainly generating significant free cash flow.

Everyone has seen that chart: semiconductor cash flow trending up, hyperscaler free cash flow trending down. But you're missing those private companies. OpenAI's and Anthropic's cash flows aren't on that chart. I also think the chart misses something important: in 2024 and 2025, even the most bullish person thought GPU rental prices would decline slowly. Bears thought they'd crash. I don't think anyone in 2024 or 2025 imagined that in 2026, prices for older GPUs would still be surging.

Chapter Two: Who Pays for the AI Buildout

Patrick O'Shaughnessy: What will the financing environment look like over the next six months? How much credit does this buildout require?

This gets into the classic capital cycle problem. If supply and demand are imbalanced, things can unravel very quickly, just like the internet bubble. If you believe hyperscalers can fund this buildout through operating cash flow, then credit tightening is less scary.

My calculation is this: The consensus assumption for hyperscalers is that they'll monetize Blackwell and Rubin compute at the rate of Ampere (chips two generations old). That's $1.3 to $1.4 trillion in hyperscaler operating cash flow. If they monetize at a rate below current Blackwell but above Ampere, it's closer to $2 trillion. That reduces credit demand by roughly $700 billion. And as installed compute reprices and operating cash flow continues to accelerate, credit metrics will improve, making financing easier.

The signals from the credit market can't be ignored. Meta issued debt last week, pricing worse than you'd expect. CDS is widening across the board. Real yields are rising. These are facts. If we need debt to fund this buildout, then yes, this is a major negative signal. But if compute reprices at current spot rates, we might not need much debt at all.

Chapter Three: GPU Spot Prices Rise, Not Fall

Patrick O'Shaughnessy: What specific data points did you hear in Silicon Valley?

I spoke with a company this morning that had rented a batch of several thousand Blackwells at around mid-$2 per GPU hour. Seven months later, for the same cluster, same scale of B200s, they're hoping to renew at under $4. That's a 50% to 60% increase in seven months.

Another inference cloud company publicly stated on a podcast that they plan to pay 100% more for Blackwell when contracts expire. This means all the hyperscalers are underreporting revenue.

There were several catalysts for July's selloff. First, Meta announced it would rent out compute, and the market interpreted this as them having oversupply and needing to cut capex. That's completely wrong. Meta saw that SpaceX had a large amount of installed compute and was selling transaction-optimized clusters to the market at prices far exceeding contract rates. Meta saw an opportunity: demonstrate high IRR on a small slice of capacity, then raise equity, then increase capex. Meta's capex telemetry data hasn't changed at all – if anything, it's become more aggressive. Right after that, they released Muse 1.1, the best model they've had in a long time.

Then Kimi K3 came out, and the market panicked about open source again. Meanwhile, Silicon Data's token index flattened. But that's because the share of open-source tokens is rising, and the way these tokens are weighted in the index creates structural flatness. A token is a token, whether it comes from a frontier model or an open-source model – the compute, memory, and electricity required to produce it are the same. What open source takes away from frontier models is profit margin, not compute demand.

Chapter Four: Claude is the Walter Cronkite of the Stock Market

Patrick O'Shaughnessy: How do you think the market is digesting this information?

There's something I've been thinking about. Mike Mauboussin has a theory that a collapse in diversity leads to bubbles and crashes. Now in the public market investing world, whether retail or institutional, every piece of news gets fed into Claude, sometimes Claude Code or Claude Agent. Claude is probabilistic, but the way people interpret news probably doesn't vary all that much.

Claude is essentially the Walter Cronkite of the stock market. Everyone believes what it says, and trades accordingly. Claude is smart, but it isn't always right. The stock market is essentially a probabilistic Bayesian interpretation of the future. You see a news headline, it gets fed into Claude, Claude interprets it in a certain way, and a large group of people trade on it.

There's an anonymous semiconductor account called TBU that posted a chart of Japanese capacitor stocks, saying "we just ran the entire capacitor cycle in 6 weeks." The stock doubled, tripled, quadrupled, then crashed. Fundamentals haven't even shown up yet, and you've already run a cycle that normally takes three years in six weeks.

Chapter Five: What Breaks the Thesis

Patrick O'Shaughnessy: If forced to identify a scenario that genuinely scares you, what would it be?

Operating cash flow ceasing to accelerate – that would be the core negative signal. That largely depends on the overall performance of Anthropic, OpenAI, Grok, Cursor, xAI, and open source.

If there's a sustained, significant drop in GPU spot prices, that would also be scary. But have you heard anyone say they have too many GPUs? Not a single one. In fact, it's quite the opposite – it sounds more like some kind of black market.

On the technical side, I think the most interesting potential risk is continual learning and sample-efficient learning. If those get solved, it could mean a temporary dip in training demand. You go from training models on 300 trillion tokens to training on just 10 trillion and then letting them learn sample-efficiently in the world – that's not great for training demand. But training's share of semiconductor demand is approaching a small, non-zero number; inference is the real driver. SSI says they're releasing a model in August, and there's a batch of new labs focused on this direction. It's great for the world, but it's hard to say whether it's positive or negative for infrastructure demand.

Chapter Six: The Game Theory of the Memory Supply Chain

Patrick O'Shaughnessy: Everyone is talking about LTAs. Can you elaborate?

We need to shift from "squeezing short-term numbers" to "exchanging durability through long-term agreements (LTAs)." The customer pays an upfront fee, with a price floor and ceiling. It's like the "labor hoarding" we talked about a few years ago when companies didn't want to do layoffs.

Let's think about the game theory of tearing up an LTA. There are four companies that truly matter at scale: Amazon's Trainium, Google's TPU, AMD, and NVIDIA, which is bigger than the other three combined. Suppose in 2027 or 2028, you're a bit tempted to tear up an LTA to get a lower price. But your market share over the next few years is essentially determined by supply chain allocation and your pre-purchased volume. If you tear up the LTA, and then leverage shifts back to the memory manufacturers, you're out.

What happens if Google tears up an LTA? It probably means oversupply, prices are falling, and capacity naturally contracts. Then this cyclical industry shifts from oversupply to undersupply. At that point, how do you think memory manufacturers will allocate Google's volume? You could destroy your entire business and brand.

It wasn't like this before. Apple was the biggest buyer and could do whatever they wanted because no one could replace their volume. But this time is different – you have at least four buyers competing, plus a bunch of startups. You tear up the LTA, and the memory manufacturer can say, "Fine, you broke the price agreement, so